Crash course: MIDI is a serial protocol for encoding of musical control signals at ~32k baud, for an effective timing resolution of ~1 millisecond. Most signals are 7-bit, yielding 128 possible note or control values, freely distributable across 16 channels. There are 120 possible control assignments; the other 7 (128 - 120 - note) are mostly reserved for meta-control, allowing private realtime parameters with up to 14-bit resolution, or bidirectional non-realtime block data transfer for firmware updates and configuration backup.
About 2/3 of the control assignments are standardized, and any MIDI device is supposed to offer a 'general MIDI' configuration with standardized sounds (trumpet, cello, clarinet etc.) controlled by standardized control assignments (cc#32 is always Volume, cc#74 is brightness aka treble, etc.). With these you can specify pretty much anything that could be written in standardized music notation in much the same manner as a music box or player piano. It is supposed to deliver a reasonable approximation of all the instruments in an orchestra and allow control of the way they're actually played, giving you up to 16 'virtual musicians' playing up to 24 notes at once. Of course, this is way too good to be true.
The hideous sounds found buried in the control panel or on Geocities-type web pages are coming out of the cheap-ass General MIDI 'synthesizer' implemented in windows and/or on your soundcard. Nobody in the real world uses this except for those awful 'home keyboards' - usually cheap, although there are also vastly overpriced versions aimed at rural churches and community centers, where people who live far away from a music store learned to play on a cheap Casio and want something that sounds better but works the same way - Bank 0, program 47 will always give you a harp sound, changing controller # 74 will always let you 'damp' the strings etc. This is a huge market which subsidizes the development of more interesting instruments. There is a parallel sub-market devoted to painstaking digital reproductions of acoustic and electric organs & pianos for more serious musicians, but which fit into a car instead of a truck.
Now, General MIDI sounds awful for three reasons: poorly sampled source sounds, crudely implemented modulation of same, and lazy transcription of musical data. The sounds are typically short (50 ms) recordings originally compressed to fit on a 16k EPROM. The modulations are usually bad approximations of inappropriately linear DSP transforms designed to run on some 8-bit uController. The transcriptions are generally bare minimum with no attention paid to articulation or the ebb and flow of musical performance. After all, it's going to sound pretty awful no matter how much you fine tune the control data, so why polish a turd?
In fairness, MIDI has been around as a protocol since 1980 and General MIDI was defined in 1991. The specification came before there were any commercial products capable of implementing it, hence the cut corners described above in a rush to look standards-compliant. Although the technology of audio reproduction has improved enormously and allows effectively perfect fidelity since then, the 'bare minimum' ethic required for standards compliance has never gone away and you can buy quite expensive keyboards that make awful-sounding GM 'music,' identical to that of 10+ years ago. It is this ugliness and sterility that Lanier objects to: far from being a highest common factor, General MIDI instead became a lowest common denominator, and is only ever heard cranking out robotic versions of jaded pop hits.
Where Lanier goes wrong is in assuming that this is a limitation of the protocol. And with only 7 bits for any given musical gesture, it is quite limited. But the main reason 'MIDI sounds so bad' is the ubiquity of the General MIDI standard. You might as well say a guitar has 21 frets and 6 strings, so it can't produce more than 126 sounds, or that your computer keyboard's 104 keys impose an unacceptable limitation on what you can write.
The reality is that while you'll never get a pleasing rendition of 16 musicians playing 24 notes from a single MIDI device, You can get astonishingly good renditions of a single musician's performance. Pretty much any real-world instrument can be beautifully synthesized and/or sampled now, and the audible changes from musical modulation accurately reproduced. Though they may use a grand piano on stage, most pro musicians will (secretly) admit they are perfectly happy with the digital version they have at home, and its multi-gigabyte sample library of conservatory pianos that are not for sale at any price. The latency and timing limitations of a ~32k serial protocol are so minute that are they below the threshold of perception for most people; they are orders of magnitude below the fastest muscular repetition rates (note that the much higher sensitivity of acoustic discrimination is for audio rather than control signals).
Latency and inconsistencies can be further reduced in importance both technically (on-the-fly reserialization yields ~100 uSec accuracy) and/or musically-aware error correction: for example, quickly strumming a guitar necessarily limits the variations in force with which each string is plucked, such that the 7-bit data for this parameter can be scaled to an appropriate range to yield a much finer expressive resolution than if it were limited to absolute values. More importantly, a quality electronic instrument allows incoming MIDI parameter data to be freely filtered, assigned and scaled to DSP parametric data running at audio speeds, from where the sky is pretty much the limit, whether your goal is the accurate reproduction of natural sounds or the creation of entirely new expressive posibilities. The main barrier to more musical expressivity is not the protocol but the physical hardware of piano keys, knobs, pedals and so on; most devices require everything to done via the fingers or tapping one's foot, and the market for more exotic sensors is small enough that they've been expensive. So MIDI on one's phone offers various new (-ly affordable) possibilities, such as waving it through the air like a conductor's baton and sending the orientation and acceleration data out via MIDI, or using the camera to track changes and reparameterizing them as MIDI. This is not to say that acoustic instruments and musicianship are any less interesting than they ever were, but to dispute the idea that music made with (or even by) computers is necessarily mechanical or musically restricted.
Now there are other standards for digital notation and reproduction, from CSound (infinite accuracy, utterly unmusical) to OSC (Open Sound Control - great system, but more for people who like synthesizer construction than for performance, and hence ignored by most instrument manufacturers). But the expressive possibilities of MIDI are so great that 30 years after its inception it remains the standard protocol. If you don't need things like the 'standard library' of General MIDI - and most serious musicians don't - then version 1.0 of the standard is still the norm, and considered perfectly acceptable even on very expensive gear (indeed, the more expensive the more likely it is to have the base protocol implementation than any 'improvements').
Incidentally, JL is also an expert on ancient instruments, and skilled in playing them. In this video, he introduces (among other things) the Khaen, an organ-like instrument from Laos and provides a neat explanation of why organs are the direct intellectual forebears of computers: http://www.youtube.com/watch?v=XW1BBbvrEYA&feature=relat...
I imagine Don Knuth's fetish for pipe organs is motivated by a similar perception.